41:21Jack Roberts · Tutorial
Every Hermes Concept Explained for Normal People
A 41-minute field guide to the open-source AI agent framework — 21 concepts, zero jargon, one tutorial that starts from zero.
May 29thThe argument in one line.
The leverage in AI agents is not asking better questions — it is building repeatable loops, scoped workspaces, and event-driven triggers so the system does the work whether you are thinking about it or not.
Read if. Skip if.
READ IF YOU ARE…
- You have installed Hermes (or a similar local AI agent) and are using it as a chat assistant rather than a system that runs independently.
- You find yourself manually prompting your agent for the same decisions — content ideas, lead research, competitor scans — multiple times a week.
- You want to understand when to use crons vs. webhooks and how to structure sub-agents without making the setup fragile.
- You are comfortable with Telegram and Notion but have not yet connected them as triggers to an agent workflow.
SKIP IF…
- You have never used Hermes or a similar local agent — you will want a getting-started tutorial before this one.
- You are looking for a passive automation setup with no code or configuration — this assumes willingness to build workflows manually first.
The full version, fast.
Most people get stuck using AI agents as reactive Q&A tools. The real upgrade is a five-layer architecture: turn every repeated decision into a scheduled loop; separate your work into Telegram topic workspaces so context stays clean; assign specialist sub-agents to specific jobs instead of overloading one; replace manual prompting with cron schedules and webhook triggers tied to your existing tools; and build a web-based mission-control dashboard so you can see, approve, and command the system from anywhere. Each layer multiplies the one before it.
Free for members
Chat with this breakdown — free.
Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.
Create a free account →Where the time goes.
00:00 – 00:51
01 · The gap
Cold open establishing the promise/install disconnect that the video will close.
00:52 – 05:11
02 · Mistake 1 — Not building loops
One-off prompts disappear; the upgrade is turning repeated decisions into scheduled workflows with sources, rules, and a cron.
05:12 – 08:34
03 · Mistake 2 — Everything in one chat
Messy context degrades output; Telegram topics as scoped workspaces with per-topic operating rules.
08:35 – 13:03
04 · Mistake 3 — One agent for every job
Overloaded agents produce overloaded context; specialist sub-agents (Nova, Sage) report to a coordinator.
13:04 – 16:14
05 · Mistake 4 — You are the trigger
Crons handle time; webhooks handle events. Stop manually prompting and let your tools trigger the agent.
16:15 – 19:00
06 · Mistake 5 — Chat as the only interface
Chat hides what is running; a web mission-control dashboard gives visibility, approval, and mobile access.
19:01 – 21:19
07 · Build order and close
Prescriptive six-step order for anyone starting from scratch; call to action for Nova/Sage links.
Lines worth screenshotting.
- Asking your AI agent a question once gives you an answer that disappears; turning that same question into a loop gives you a system that compounds.
- A prompt becomes a workflow the moment the agent is checking sources, applying rules, and returning something you can act on without extra research.
- Do the task manually first — every source it needs, every rule it should follow — before scheduling it as a cron.
- Messy context is the enemy of agent quality; the more capable the agent, the more damage a polluted context window does.
- Telegram topics are free, zero-code workspaces: each topic gives the agent a different mode, memory, and set of tools with no extra infrastructure.
- More agents does not mean better output; more focused agents with narrow lanes and specific tools means better output.
- You do not want your YouTube agent thinking about your grocery list when it is generating video ideas.
- Webhooks turn your existing tools into triggers: a Notion card moved to a column, a form submitted, a PR opened — any state change becomes a command.
- A polling cron closes the gap when an app has no webhook: check every few minutes, compare what changed, act only when the condition is met.
- Making yourself the trigger for everything means the system only works when you are paying attention to it.
- Chat is good for commands and bad for visibility — a mission-control dashboard gives you the confidence to let the agent do more.
- A toy setup hides everything in chat; a real setup gives you command, visibility, and approval.
- The build order that minimizes wasted effort: one loop, one workspace, one specialist agent, one cron, one event trigger, then mission control.
- Teaching the agent how your work runs is more valuable than teaching it what your work is.
Five layers that turn an AI agent into a system.
WHAT TO LEARN
A single AI agent answering one-off questions is useful but not compounding — the real leverage starts when you layer loops, scoped workspaces, specialist sub-agents, event triggers, and a visibility dashboard on top of each other.
02Mistake 1 — Not building loops
- Any decision you make more than once a week is a candidate for a repeatable loop — define the sources, the rules, and the output format, then let the cron run it on a schedule.
- Validate every workflow manually before scheduling it: confirm each data source returns clean results and each rule fires as expected before handing control to automation.
03Mistake 2 — Everything in one chat
- Separate workspaces by job type, not by project — one workspace for content, one for admin, one for main work — and give each workspace a simple operating rule the agent follows automatically.
04Mistake 3 — One agent for every job
- A coordinator agent that dispatches to specialist sub-agents and reads their reports will consistently outperform a single generalist agent given the same task.
05Mistake 4 — You are the trigger
- Crons handle time-based recurrence; webhooks handle event-based activation. When a tool has no webhook, a polling cron checking every few minutes is a workable substitute.
- The moment you stop being the trigger — letting a Notion card move, a form submit, or an email arrive activate your agent automatically — the system starts producing value while you are offline.
06Mistake 5 — Chat as the only interface
- A chat interface cannot show you what is running, what failed, or what needs approval at scale; a simple web dashboard with authentication solves all three problems and makes the agent worth trusting with more work.
Terms worth knowing.
- Hermes Agent
- A local AI agent application (by Nous Research) that can run background tasks, connect to external tools via APIs, and be accessed through interfaces like Telegram.
- Cron / cron job
- A task scheduled to run automatically on a time-based interval (e.g., every morning, every Friday). The term comes from the Unix cron scheduler.
- Webhook
- An event-driven trigger: instead of running on a schedule, a webhook fires a workflow the moment a specific thing happens in an external app (a form submit, a status change, a file upload).
- Polling cron
- A scheduled job that periodically checks whether a condition has changed and acts only when it has — a workaround when an app does not support native webhooks.
- Sub-agent
- A specialized agent with a narrowly defined job, its own tools and memory, that reports results back to a coordinator agent rather than handling everything itself.
- Mission control
- A web dashboard that surfaces what an agent system is doing — running jobs, completed tasks, failures — so the operator can see, approve, and command without relying on chat alone.
- Telegram topics
- Thread-style sub-channels inside a Telegram group that act as separate workspaces, each carrying its own context and operating rules for the agent.
- Nova
- An open-source specialist sub-agent built for YouTube research — checking trends, competitor outliers, and content pipeline — released by the video's creator.
- Sage
- An open-source specialist sub-agent built for X (Twitter) content strategy, parallel in structure to Nova but scoped to social content rather than YouTube.
Things they pointed at.
Lines you could clip.
02:21
“Do not ask what can Hermes answer for me. Ask what decisions do I keep making manually.”
04:52
“The value is not in having Hermes do the task once. The value is that it can turn the task — that grunt work you find yourself doing every day — into a repeatable operating system.”
07:28
“The more powerful your agent becomes, the more dangerous messy context becomes.”
10:56
“More agents does not automatically mean better. More focused agents means better.”
16:24
“Stop making yourself the trigger for Hermes to wait for you to ask for things and let your tools, your emails, your data trigger Hermes.”
18:48
“A toy setup hides everything in chat. A real setup gives you command, visibility, and approval.”
20:53
“Do not just ask Hermes to help. Teach it how your work runs. That is the leverage.”
Word for word.
Read-along
Don't just watch it. Burn it in.
See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.
metaphoranalogy
00:00You see people online using Hermes to run their business. It finds them leads.
00:06It writes briefs. It watches the Internet, sends them updates on their phone, runs little agents in the background, then you install it and ask it to summarize a PDF.
00:19That gap is what this video is about. So after a hundred days with Hermes, I wanna show you actual bridge. How you go from asking Hermes random questions to building repeatable loops, telegram workspaces, sub agent teams, event triggers, and a mission control dashboard you can open from anywhere in the world by turning one boring task at a time into a system that runs without you.
00:48Let's get started. The first mistake that I've learned throughout my first one hundred days is using Hermes like a smarter search box, Asking it for things like, uh, hey, Hermes, give me ideas, summarize this, research that, write a draft about this.
01:08I mean, that is pretty useful and it's a big massive major reason for AI agents being useful. But that or those types of tasks don't compound because essentially you get an answer to your question but the work disappears. The better move is to turn the task into a repeatable loop.
01:30Here's the difference. If I say give me YouTube ideas, Hermes can give me YouTube ideas. But if I say every time I ask you to give me YouTube ideas, check my posted videos, check my rejected ideas, check my Notion content board, check current demand, check competitor outliers, then tell me the one video I should film next, why it matters now, what proof I need, and what the first thirty seconds should show.
02:03See, that's not a prompt anymore. That's an entire workflow. Hermes is checking for sources.
02:10It is applying rules. It's making decisions. It is returning something I can act on, something that is actionable, and that is the first real upgrade.
02:21Do not ask what can Hermes answer for me. Ask what decisions do I keep making manually because that is usually where the loop is.
02:31For me, one loop is reaction opportunities. I do not want to wake up, scroll x, scan AI news, check what is fresh, think through whether it matters, and then decide if I should post.
02:47So Hermes can do that for me every day, every couple of hours. It can check for fresh AI stories. It can check expos by me providing it with a rock API.
02:59It can check for product launches, filter for YouTube relevant topics. It can prioritize things from the last hour or skip stale topics and things that it finds are duplicates.
03:13A massive pro tip by the way, do the thing manually first. Make sure that the loop works manually first. For example, for my YouTube video ideation, I've made sure that Hermes can fetch me every single individual piece of the puzzle.
03:34I've made sure that it can use and fetch all of the information and data that it needs from the sources it requires, that I made sure that it knows how to use that data to come up with video ideas, then I made sure that the video ideas are actually good and actual signals.
03:53Once I confirmed all that, then I schedule it as a cron, which is just a fancy word for a repeatable job, something that happens on an occurrence.
04:06Every day, every couple of hours, you you choose. You can say like every morning, check what matters. Every Friday, audit video ideas.
04:15Every day, scan competitors, and cron should make sense to you. I mean, check these sources, follow these rules, ignore these things, send me this output, run it on a schedule.
04:28This is how you go from asking Hermes for help to having Hermes report back, and this works for almost any job. A founder can summarize new leads every morning. A creator can scan video opportunities every week.
04:43A developer can open PRs daily. A salesperson can find follow ups they might have forgot. The value is not in having Hermes do the task once.
04:52The value is that it can turn the task, that grunt work that you find yourself having to do every day into a repeatable operating system. But the second you start creating loops, you hit the next problem.
05:07Everything gets noisy if all the work lives in one place. My first Hermes setup was basically just .
05:18I would talk to Hermes about my YouTube ideas, my expos, my business strategy, my personal tasks, just random questions and it was all in the same chat. And granted, that can feel a lot simpler, especially when you're starting out until it becomes impossible to tell what anything was for.
05:41The fix was not a more complicated app. It was Telegram topics.
05:48I stopped treating Telegram like one conversation with an AI, and I started treating it like separate different workspaces.
05:57I mean, I have one topic that is dedicated to YouTube. I it's all I talk about with Hermes about. This is where one of my sub agents also live called Nova, but we'll get into that later.
06:09One topic is x. That one is built for me to understand what is happening around x. I mean, have I been offline the entire day?
06:17Do I need a quick refresher? Can I ask Hermes to fetch sources for one of my tweets that I'm writing or putting out? Another topic is general, and I strongly recommend you have a general topic where it's a sort of catch all topic that you can ask any question that you have that doesn't really fit into your other topics.
06:36One topic is react, which sends me reaction opportunities for things I can talk about on x.
06:43This is especially helpful for me if I've been offline or on meetings for an entire day and have not had time to see what is happening with the AI market. The magic is not Telegram or even Hermes.
06:57The magic is that every workspace starts having a job. Every Telegram topic is a completely separate workspace.
07:05When I message Hermes in the YouTube topic, Hermes does not need to guess what mode it's in. It doesn't need to load a specific context and store away another type of context.
07:18It is already in the YouTube room, and that changes the quality of the output you get immediately. Because the more powerful your agent becomes, the more dangerous messy context becomes.
07:32If I'm asking for a video idea, I do not want the agent thinking about my grocery list, a code bug, and some random email from yesterday. I want it inside the content workspace using the content rules and content agents or sub agents.
07:50That is the practical setup I would recommend. Start with three topics. One for your main work, one for content or research, and one for admin administrative stuff.
08:05Then give each topic a simple operating rule. For content, the rule might be before suggesting anything, check what I already posted, what I rejected, what is currently trending, and whether I can actually make it.
08:20For admin, the rule might be organize the task, draft the response, but ask before sending anything externally. Now Hermes is not just answering messages, it is working inside the right room.
08:35That is the second upgrade. Chats become workspaces. But workspaces only solve where the work happens.
08:44They do not solve who does the work. That is where sub agents changed the setup for me. Let's go to mistake number three.
08:54And I know one Hermes agent can do a lot and that is the trap.
09:01Because when something can do a lot, you start giving it everything. Research the market, check competitors, inspect Notion, find title ideas, write the hook, summarize the hundreds of decisions that I have to make every day.
09:17I mean, it can do that, but it will get very messy very quickly and that's what happened with me. I noticed that the better pattern is sub agents.
09:28Think of the main agent as being the main operator, the one who coordinates with all of your other employees.
09:39It does not need to personally inspect every detail. It needs clean reports from specialists, then it needs to make the final call.
09:48For example, if I want Hermes to pick my next video, I can split the work. One sub agent checks my posted and filmed pipeline.
09:58One sub agent checks competitor outliers. One sub agent checks current demand. The main agent reads the reports and decides what I should actually film.
10:10Now the work is a lot cleaner, the context is cleaner, the decision ends up also being a 100 times better. Just try it and you'll see.
10:19And this is where you can use hack number one in this hack number three because once you define how that workflow looks like, then you can turn it into a repeatable cron or a repeatable skill.
10:35Meaning every time you ask for say a YouTube idea, it doesn't have to guess and start from scratch. It already knows, alright, let's deploy our three sub agents. One does this, one does that, one does that, so on and so forth.
10:50You get the idea. However, this is also where a lot of people get agents wrong.
10:56More agents does not automatically mean better.
11:01More focused agents means better. Each agent needs a specific lane.
11:08I love to give this analogy, but you don't want your plumber to be your dentist or maybe he's a very talented plumber who also have or happened to take a degree in dentistry.
11:20Uh, I mean, hey, don't judge him.
11:25But that is why I built highly specialized agents like Nova and Sage.
11:31Nova is my YouTube agent. She has a wealth of knowledge and skills. She's trained on finding and checking YouTube trends, comparing video outliers, comparing competitor posts, and finding the best video ideas and helping me through my research, ideation, everything I need for YouTube.
11:51And I've open sourced her, meaning you can download and install her for free. I'll leave the link in the description. And Sage is my x and content strategy agent, which is basically the same thing as Nova but dedicated for x.
12:07I've also open sourced him and he is free to download and install, and I'll include the link in the description for both. Hooray.
12:17You can copy their structure or turn them into your own specialized agents, but the point is not my exact agents. The point is the pattern.
12:28Your YouTube agent should not be the same as your coding agent and your coding agent should not be the same as your personal admin agent and your plumbing agent should not be the same as your dentist agent. Anyway, starting to feel like I'm going in circles around here.
12:43Different job means different tools, mean different memories, means different permissions. The list goes on and on.
12:50That is how you stop building one overpowered assistant and start building a team. And once you have a team, the next question becomes obvious.
13:00How does the team know when to act? This is where event triggers come in. Cron handles time.
13:09Webhooks handle events. I know what you're thinking, what is he talking about? That sounds way too technical, but let me explain this to you because the idea is simple.
13:20Crons are repeatable events as we discussed. Things like do this every single morning but a webhook says do this when something happens.
13:32That is the difference between cron's and webhooks. Cron's repeat something after every passing of a certain amount of time and webhooks do something when something happens.
13:43And listen to me because the second one, webhooks, is where Hermes starts to feel alive. Here's an example. I have a Notion content board.
13:53When I move a video idea to film tab, that movement triggers work. It triggers a webhook.
14:01The second a video moves to to film, Hermes can validate it. It starts researching it that I already post something that is too similar.
14:11Is there current demand for that video idea that I just moved? What is the strongest title for that video? Should I film this now, later, or just kill it, abandon it?
14:22That is the difference between a database and a control service. Notion is not just where the ideas live. .
14:32It becomes the trigger to a chain of events and work. And you can use this yourself anywhere you want. Uh, let me give you some ideas.
14:42A form gets submitted. Hermes researches the lead. Uh, GitHub PR opens.
14:48Hermes reviews the risk. A customer pays, Hermes adds it to the daily summary. Uh, a file gets uploaded, Hermes extracts the useful parts.
14:59You receive an email, Hermes reads it and summarizes it to you and comes up with a draft of what you could reply to. The setup is pretty simple.
15:10You can get as wild as you want with the ideas. All you need to do is create a webhook route in Hermes and super simple by the way. Don't let this intimidate you.
15:21You can literally ask your Hermes, Hermes, how do we create webhook triggers? And it will guide you. Or it will do it for you if you have it connected to like a super smart model, it will connect to the app itself.
15:36Uh, the only thing you'll need to do on your end is, you know, tell it what you need from it. And then whenever that event happens, Hermes runs the workflow every single time. And if the app does not support a clean webhook, you can still get close with a polling cron.
15:54What in the world is a polling cron? It's a cron that keeps pulling. Did this event happen every five minutes?
16:01It didn't? Okay. Never mind.
16:02I won't do anything. Oh, now it did in the next five minutes. Oh, it just did.
16:07Alright. Let's run the Cron. Or simply put, Hermes can check every few minutes, compare what changed, and only act when the right thing happens.
16:16The bigger lesson is this. Stop making yourself the trigger for Hermes to wait for you to ask for things and .
16:29That is when your system starts working even when you are not thinking about it, even when you're not working. But once you have loops, topics, sub agents, and triggers, you run into the final problem.
16:43You cannot manage a real system from a chatbot forever. You need a place to see the machine.
16:52On to mistake number five. Chat is good for commands. Chat is bad for visibility.
17:00If Hermes is doing real work, I need to know what is happening, what is running, what finished, what failed, where are the different things that I need and ask for in order to be productive on a day to day basis.
17:16This is why I highly recommend that you build yourself a mission control. Mission control is the dashboard layer.
17:25It's your information layer. Think of you opening Jarvis and you see all of this information and you're like, wow, this is so cool and futuristic.
17:34This is your mission control. For me, it is a live Vercel website connected to my workflows.
17:41Uh, I have set up Google authentication, Gmail authentication so that I can log in to it through my Gmail.
17:50That way, you know, some random person can't just find the website link and log in to himself and control my entire life or ruin my entire life.
18:00And what this also means is I can open it on my phone. I can log in and I can control Hermes straight from my phone.
18:10I can have every single thing that I need right on my phone. And the great thing about that is I can start using it anywhere I am in the world.
18:20I don't need to be next to my local machine in case you're running Hermes on a local instance. I don't need to be on my desktop if I'm out in the road traveling at the gym, I need something done, or I'm on holiday and I don't have my work equipment with me, I can still be productive by literally asking Hermes to do the thing for me.
18:41Mission controls give you the confidence to let Hermes do more because you can actually see what it is doing. That is the difference between a toy setup and a real setup. A toy setup hides everything in chat.
18:54A real setup gives you command, visibility, and approval. And if I were to download and install Hermes from scratch, I would build it in this order.
19:06One boring loop, one telegram workspace, one specialist agent, one scheduled cron, and one event trigger.
19:15Then mission control once the system gets too big for just chat. That is a great path for you to level up your Hermes skills and control. Answers become loops, chats become workspaces, one agent becomes many agents and a team, manual work becomes triggers and invisible work becomes your mission control.
19:40That is what one hundred days with Hermes taught me. The point is not to use every feature. The point is to transfer your workflow into the system.
19:51That is when Hermes stops feeling like a chatbot, and that is when it becomes an operating layer to your entire life and a genuine point of improvement to your quality of life.
20:04So if you wanna start, don't overbuild. Pick one task you already repeat over and over again and turn it into a loop.
20:13Run it through Hermes. Ask it to help you. Then from there, add one layer at a time.
20:19Add workspaces, then once you've done that, add sub agents, then crons, then webhooks, then mission control.
20:27And if you want the specialist agents I showed, Sage and Nova, those are open source on my GitHub. I'll include the links below.
20:36If you want the template to my mission control, I'll also include the link to that below. It's also open source and free for you to download. Feel free to copy them, modify them, or use them as templates.
20:49Uh, but the real lesson here is bigger than just those agents. Do not just ask Hermes to help. Teach it how your work runs.
20:59That is the leverage. And if you've enjoyed this video, make sure to subscribe because I've got a ton more like it on my channel and a ton more coming your way. Oh, would you look at that?
21:11The YouTube algorithm gods really think you'll enjoy this video. So I'll see you there.
The Hook
The bait, then the rug-pull.
There is a gap between what Hermes looks like on someone else's screen and what it does the first time you install it. This video is the bridge — built from 100 days of daily use — that closes that gap with five concrete mistakes and the system that fixes each one.
Named ideas worth stealing.
00:19list
The Five-Layer Agent Stack
- Repeatable loops
- Telegram workspaces
- Specialist sub-agents
- Crons + webhooks
- Mission control
Each layer is additive: loops provide the value unit, workspaces keep context clean, sub-agents divide labor, triggers automate activation, and mission control makes it all visible.
Steal forAny AI agent setup regardless of tool — the pattern works with Claude, GPT, or any orchestration layer
03:13concept
Do It Manually First
Before scheduling any workflow as a cron, manually confirm every data source, rule, and output is working correctly. Only automate what you have already validated.
Steal forReducing automation debt — prevents scheduling broken workflows that silently fail
13:07concept
Cron vs. Webhook decision rule
Crons handle time-based recurrence (every morning, every Friday). Webhooks handle event-based activation (when X happens). When a tool lacks webhooks, a polling cron checks for the condition every N minutes.
Steal forAny automation architecture decision — the framing is tool-agnostic
How they asked for the click.
VERBAL ASK
19:01subscribe
“If you enjoyed this video, make sure to subscribe because I have got a ton more like it on my channel.”
Soft ask at the very end after the full content payload; also links Nova, Sage, and Mission Control template in description throughout.
MENTIONED ON CAMERA
FROM THE DESCRIPTION
OTHER LINKSAlso linked in the description.
Visual structure at a glance.
